vector search api

Hybrid search using full-text and vector search

• This is a search function which used llama LLM model to generate embeddings (1024)
• Embeddings are stored in posgres leveraging pgvector library
• REST endpoint written in ExpressJS

Setup Instructions

Follow these steps to set up and run the project:

1. Install PostgreSQL

   • Download and install Postgres.app with PostgreSQL 16 from Postgres App.
   • Open Postgres.app, initialize PostgreSQL, and go to server settings.
   • Select the user with your system name and change the password to admin.
   • Start the PostgreSQL server.

2. Configure the Project

   • Navigate to the config folder in the project directory.
   • Open db.js and update line 3:
     • Change the username and database name from mayanksharma to your system username.

3. Set Up the Database

   • In the Postgres app, double-click on the database with your username to open a terminal.
   • Run the following command in the terminal:

     CREATE EXTENSION vector;

4. Install Ollama

   • Download and install Ollama from Ollama Download.
   • After installation, do not run any model as prompted.
   • Open a terminal or command prompt and run:

     ollama pull snowflake-arctic-embed

   • Again, do not run any model as prompted after the installation.

5. Install Project Dependencies

   • In the project terminal, run the following commands to install dependencies and start the server:

     npm install
     node server.js

6. Install REST Client Extension

   • Download and install the "REST Client" extension (blue icon) for your code editor.

7. Test the API

   • In the root directory of the project, open the api.http file to test the API endpoints.

Additional Resources

• Postgres App
• Ollama Download

---

API documentation and usage examples

• POST: /api/v1/magazine/hybridsearch/[page_number] Returns the hybrid search results
  • 1. Search endpoint content-type: application/json

       {
          "query": "your_search_query"
       }

Only for adding data, not a part of task submission

• POST: /api/v1/magazine Add magazine endpoint
  • 1. Search endpoint content-type: application/json

       {
          "title": "magazine_title",
          "author": "author_name",
          "category": "magazine_category",
          "content": "magazine_content"
        }

---

Performance Report

Performance considerations

I have used PostgreSQL with pgvector (storing embedding vectors) and tsvector (storing content text).

Requirement: search from 1 million records

• Added Hierarchical Navigable Small Worlds (HNSW) indexes for vector search on content embeddings Reason: Search requires high recall, which makes hnsw better than ivfflat Reference

  • vector_ip_ops
  • vector_cosine_ops
  • vector_l1_ops

• Added indexes for title, author and content

  • GIN indexing is used for content in TSVECTOR datatype

• Pagination added for reduce load times

  • limit and offset in queries

• Profile: Peak

• Virtual Users: 20

• Test Duration: 5 minutes

• Endpoint hit: POST /api/v1/magazine/hybridsearch/1 ("glasgow", "game", "business", "shubham", "food" and "modern")

• Total requests sent: 10,915

• Request per second: 35.62

• Avg response time: 116 ms

---

Hybrid search implementation explained

Two individual services for text search and vector search is used

Embeddings are generated by Meta llama "snowflake-arctic-embed" model, being lightweight.

• STEP 1: Common objects from both vector and full text search results are shown first,

• STEP 2: followed by objects from only text search,

• STEP 3: rest of the objects from vector search.

---

Relevance testing for results

• query: vector "glasgow", return "Celtic feast journal" which has "Scotland written in content"

• query: vector "shortbread", returns "Celtic feast journal" as "shortbread" is related to "scotland"

• query: keyword/full-text "shubham", returns "Physics Refresher" which has author name "Shubham Thorve"

• query: keyword/full-text "mayank", returns "Digit Gaming" which has author name "Mayank Khurana"

• query: keyword/full-text "month", returns "Dalal Street Journal" which has content "All about video games this month"

Database Schema ORM is located in /model

---